[RFC PATCH] ARM: decompressor: implement autonomous KASLR offset calculation

[Kees Cook]

On Thu, Aug 17, 2017 at 12:22 PM, Ard Biesheuvel
<ard.biesheuvel at linaro.org> wrote:
> On 15 August 2017 at 21:29, Kees Cook <keescook at chromium.org> wrote:
>> On Tue, Aug 15, 2017 at 1:15 PM, Ard Biesheuvel
>> <ard.biesheuvel at linaro.org> wrote:
>>> This enables KASLR for environments that are not KASLR-aware, or only
>>> to a limited extent. The decompressor collects information about the
>>> placement of the zImage, DTB and initrd, and parses the /memory DT
>>> node and the /memreserve/s and /reserved-memory node, and combines this
>>> information to select a suitable KASLR offset, and proceeds to decompress
>>> the kernel at this offset in physical memory. It then invoked the kernel
>>> proper while passing on this information, so that it can be taken into
>>> account to create the virtual mapping.
>>>
>>> This code shuffles some registers together to create a poor man's seed,
>>> which will be superseded by the value of /chosen/kaslr-seed if present.
>>>
> [...]
>>
>> Without kaslr-seed, perhaps also include build-time entropy (as I did
>> in x86's decompressor with build_str):
>>
>> static unsigned long get_boot_seed(void)
>> {
>>         unsigned long hash = 0;
>>
>>         hash = rotate_xor(hash, build_str, sizeof(build_str));
>>         hash = rotate_xor(hash, boot_params, sizeof(*boot_params));
>>
>>         return hash;
>> }
>>
>
> Shouldn't we use something with better diffusion than rotate and xor?
> If we can afford to decompress ~10 MB worth of vmlinux, I'm sure we
> can afford to crc() the device tree blob.

The final kaslr_get_random_long() (in arch/x86/lib/kaslr.c) uses a
circular multiply. That could be done earlier, I suppose, but I wasn't
too concerned given common the availability of RDRAND on x86 has
become.

>> You're effectively starting with hash == r3. This could be further
>> enhanced with a __latent_entropy string when that plugin is enabled:
>>
>> static u8 compile_time_entropy[32] __latent_entropy;
>>
>> ...
>>
>> seed = rotate_xor(seed, compile_time_entropy, sizeof(compile_time_entropy));
>>
>> And toss in the fdt too? (I have no idea if this is the correct way to
>> do this....)
>>
>> seed = rotate_xor(seed, fdt, fdt_totalsize(fdt));
>>
>
> As mentioned by several people, this is a nice way to add
> /chosen/kaslr-seed to the mix if it is there, and still have some
> pseudo=entropy otherwise, especially given the fact (as Daniel figured
> out) that Android puts device serials and some boot timing related
> variables on the kernel command line.

Cool, yeah.

-Kees

-- 
Kees Cook
Pixel Security